On Wed, Sep 25, 2019 at 12:06 PM Martin Liška <mli...@suse.cz> wrote:
>
> Hi.
>
> Similarly to SLP pass, we should probably set TODO_update_ssa
> when a SLP BB vectorization happens from the normal vect pass.
>
> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
>
> Ready to be installed?
Hmm, this was supposed to be handled by
if (num_vectorized_loops > 0)
{
/* If we vectorized any loop only virtual SSA form needs to be updated.
??? Also while we try hard to update loop-closed SSA form we fail
to properly do this in some corner-cases (see PR56286). */
rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
return TODO_cleanup_cfg;
}
but there isn't an equivalent of num_vectorized_bbs here. Given the above
effectively short-cuts your patch (all paths only ever set ret =
TODO_cleanup_cfg
right now...) it will work without pessimizing things by running
update-ssa twice.
Can you check whether TODO_update_ssa_only_virtuals is enough as a fix?
Otherwise OK.
Thanks,
Richard.
> Thanks,
> Martin
>
> gcc/ChangeLog:
>
> 2019-09-25 Martin Liska <mli...@suse.cz>
>
> PR tree-optimization/91885
> * tree-vectorizer.c (try_vectorize_loop_1):
> Add TODO_update_ssa similarly to what slp
> pass does.
>
> gcc/testsuite/ChangeLog:
>
> 2019-09-25 Martin Liska <mli...@suse.cz>
>
> PR tree-optimization/91885
> * gcc.dg/pr91885.c: New test.
> ---
> gcc/testsuite/gcc.dg/pr91885.c | 47 ++++++++++++++++++++++++++++++++++
> gcc/tree-vectorizer.c | 2 +-
> 2 files changed, 48 insertions(+), 1 deletion(-)
> create mode 100644 gcc/testsuite/gcc.dg/pr91885.c
>
>
